Applied Numerical Mathematics 116 (2017) 230–237

Contents lists available at ScienceDirect

Applied Numerical Mathematics

www.elsevier.com/locate/apnum

Stability and boundedness of numerical approximations

to Volterra integral equations
Eleonora Messina a,∗ , Antonia Vecchio b
a
Dipartimento di Matematica e Applicazioni, Università degli Studi di Napoli “Federico II”, Via Cintia, I-80126 Napoli, Italy
b
C.N.R. National Research Council of Italy, Institute for Computational Application “Mauro Picone”, Via P. Castellino, 111, 80131 Napoli, Italy

a r t i c l e i n f o a b s t r a c t

Article history: Volterra Integral Equations (VIEs) arise in many problems of real life, as, for example,
Available online 26 January 2017 feedback control theory, population dynamics and fluid dynamics. A reliable numerical
simulation of these phenomena requires a careful analysis of the long time behavior of
Keywords:
the numerical solution. Here we develop a numerical stability theory for Direct Quadrature
Volterra Integral Equations
Numerical stability
(DQ) methods which applies to a quite general and representative class of problems. We
Direct quadrature methods obtain stability results under some conditions on the stepsize and, in particular cases,
unconditional stability for DQ methods of whatever order.
© 2017 IMACS. Published by Elsevier B.V. All rights reserved.

1. Introduction

In the mathematical representation of real life history-dependent problems (such as mechanical systems, electric circuit,
epidemiology, population growth,...), systems of Volterra integral equations are widely used (see e.g., [5,9–11]) and have
been subject of extensive investigations over the years (see [5,6,10] and the bibliography therein). Here, we are concerned
in the numerical simulations of linear VIEs of the type

t
y (t ) = f (t ) + k(t , s) y (s)ds, t ≥ 0, (1)
0

where k, defined for 0 ≤ s ≤ t < ∞, with values in the space Rd×d of d × d matrices, is a Volterra kernel (k(t , s) = 0 when
s > t) and the forcing function f is defined for t ≥ 0 and takes values in Rd . We assume that f (t ) and k(t , s) are at least
continuous for t ≥ 0 and 0 ≤ s ≤ t < +∞ respectively. Existence and uniqueness results can be found in [10].
The stability analysis of the numerical method used in the simulation of these equations is essential to assure that the
approximating solution preserves the characteristics and the long time behavior of the continuous problem.
An analysis of this kind is already well known in the literature (see [1,2,14,13,17] and [4,3] for a thorough discus-
sion), but in spite of the considerable advances that have been made in the last years, there are still many aspects to
investigate and open question to answer. The main idea in the numerical stability analysis is to identify more and more
realistic model problems (representative of the applications) and to study how numerical methods behave when simulating
them.

* Corresponding author.
E-mail addresses: eleonora.messina@unina.it (E. Messina), antonia.vecchio@cnr.it (A. Vecchio).

http://dx.doi.org/10.1016/j.apnum.2017.01.011
0168-9274/© 2017 IMACS. Published by Elsevier B.V. All rights reserved.
 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237 231

Our approach consists in discretizing some suitable classes of equations (1) by Direct Quadrature (DQ) methods and
study the asymptotic behavior of the numerical error as a function of the stepsize h.
In Section 2 we introduce some basic material needed in the paper. In Section 3 we define the model problem and
obtain conditions on the stepsizes h for the DQ methods to be stable. Furthermore, we identify subclasses of problems
for which the methods turn out to be unconditionally stable. In Section 4 we prove the boundedness of the numerical
solution, whenever the forcing term f (t ) is bounded. In Section 5 numerical examples are given and Section 6 contains
some concluding remarks.

2. Preliminaries

Consider the general DQ–(ρ , σ ) method for the numerical integration of the linear VIE (1)

n 0 −1
 
n
yn = f n + h w nj knj y j + h ωn− j knj y j , n = n0 , n0 + 1, . . . , (2)
j =0 j =n0

where yn  y (tn ), f n = f (tn ), knj = k(tn , t j ), with tn = nh for n = 0, 1, . . ., h > 0 is the stepsize and w nj , ω j are the weights
of a Gregory quadrature formula (which corresponds to Adams–Moulton (ρ , σ )-quadrature formula, see [4,13,18] for a
general discussion on DQ–(ρ , σ ) methods). Here, we assume that the weights are non-negative and that y 0 = f 0 , . . . , yn0 −1 ,
n0 ≥ 1, are given starting values. The property

W = sup ωn < +∞ (3)

n

of (ρ , σ ) weights (see for example [18]) will be useful in the following discussion.
Let us write the error equation for the DQ method (see for example [12, p. 104])

n 0 −1
 
n
en = T n (h) + h w nj knj η j + h ωn− j knj e j , n = n0 , n0 + 1, . . . (4)
j =0 j =n0

where en = y (tn ) − yn is the global error, T n (h) represents the local discretization error and η j , j = 0, . . . , n0 − 1 are the
starting errors. According to [18] the local error T n (h) has an expansion of order p, that is (see definition in [12, p. 105])
there exists a function Q (t ) such that

T n (h) = Q (tn )h p + R n (h)hq , q > p . (5)

From now on we assume that supn  Q (tn ) ≤ Q̄ , with Q̄ independent of n and h and sup0<h<h∗ supn≥0  R n (h) is finite for
small h∗ > 0. A remark on these assumptions is reported at the end of Section 3.
All the analysis below will be based on the following definition of stability of the numerical scheme (2):

Definition 1. The numerical method (2) for solving a class of Volterra integral equations (1) is said to be stable for that class
of equations if, using a given stepsize h > 0, the global error en given by (4)–(5) satisfies

en  ≤ Eh p , ∀n ≥ 0,
with E constant independent of n and h.

According to Definition 1, our request for a numerical method to be stable is not only that the local errors do not build
up catastrophically in the step by step process, but also that the global errors remain bounded for a fixed stepsize h > 0,
and may be made arbitrary small reducing h, over long-time intervals.
From now on we will assume that h satisfies

det( I − hω0 knn ) = 0, (6)

where I is the identity matrix of size d, in order to assure the solvability of the discrete problem (2). Furthermore, we
assume the starting errors η j , j = 0, . . . , n0 − 1, to be accurate at least of order p, so to incorporate them into the local
n0 −1
error T n (h) (that is T n (h) is now equal to T n (h) + h j =0
w nj knj η j ) and make (5) still valid.

3. Numerical stability for linear equations

This section is devoted to the study of the behavior of a numerical method of the form (2) when applied to problem (1).
The class of problems that we are going to examine consists of equations (1) where the kernel k satisfies:
232 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237

a) K = supt ≥s k(t , s) < ∞, ∀t ≥ s ≥ 0,

t
b) sup k(t , s)ds ≤ α < 1, and
t ≥0
0
t  
 ∂ k(t , s) 
 
c ) sup  ∂ s  ds < +∞.
t ≥0
0

Remark 1. It is important to point out that the class of problems defined above is a standard one in the context of stable
VIEs. As a matter of fact, the results known in the literature on the global behavior of the solutions are usually based on
hypothesis b) on the kernel k (see for example [10, sec. 9.8 and 9.9]).

The following lemma represents the link between the continuous and discrete classes of problems considered here.

Lemma 1. Assume that, for the kernel k(t , s) in equation (1), a), b) and c ) hold and let
⎛ ⎞

n0  t  
 ∂ k(t , s) 
A = ⎝K (ω j − ω̃ j ) + W sup   ⎠
 ∂ s  ds , (7)
t ≥0
j =0 0

where ω j are the convolution weights of the DQ method (2), ω̃ j are defined as

ω j if ω j ≤ 1
ω̃ j = , j = 0, . . . , n 0 , (8)
1 if ω j > 1

and W is given in (3). Then,


n
sup h ωn− j knj  ≤ α + Ah, (9)
n≥n0
j =n0

where α is given in b) and h is the stepsize.

Proof. Consider


t j +1

t j +1

t j +1

k(tn , s)ds = hk(tn , t j ) + k(tn , s)ds − k(tn , t j )ds

tj tj tj

 
t j +1 s
∂k
= hk(tn , t j ) + (tn , x)dxds.
∂x
tj tj

So


n 
n 
t j +1

h ωn− j knj  ≤ ωn− j k(tn , s)ds

j =n0 j =n0 tj
(10)
t j +1 

n 
 ∂k 
+h ωn− j  (tn , s) ds.
 ∂s 
j =n0 tj


n 
t j +1

Let us consider ωn− j k(tn , s)ds, and observe that ωi = 1 for i > n0 (it comes from [4, (2.6.16’), p. 72] and from the
j =n0 tj
consistency conditions of Adams methods for ODEs). Then, for n > 2n0 ,
 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237 233


n 
t j +1
n−n0 −1 j +1

t

n
ωn− j k(tn , s)ds = k(tn , s)ds + ω̃n− j
j =n0 tj j =n0 tj j =n−n0
(11)

t j +1

n 
t j +1

k(tn , s)ds + (ωn− j − ω̃n− j ) k(tn , s)ds.

tj j =n−n0 tj

Hence, from (10) and (11)


n tn 
n
h ωn− j knj  ≤ k(tn , s)ds + (ωn− j − ω̃n− j )
j =n0 0 j =n−n0

t j +1 

t j +1

n 
 ∂ k(tn , s) 
k(tn , s)ds + h ωn− j  
 ∂ s  ds.
tj j =n0 tj

Since a) is true, we obtain


n tn
h ωn− j knj  ≤ sup k(tn , s)ds
n≥n0
j =n0 0
⎛ tn 
⎞

n0  
 ∂ k(tn , s) 
⎝
+h K (ω j − ω̃ j ) + W sup   ⎠
n≥n0
 ∂ s  ds . 2
j =0 0

Since α < 1 for hypothesis b), there exists h∗ > 0 such that
α + Ah < 1, ∀h ∈ (0, h∗ ], (12)
with A defined in (7). Then, for the error defined in equation (4) the following stability result is straightforward.

Theorem 2. Assume that a), b) and c ) hold for equation (1) then

Q̄ h p
supn≥n0 en  ≤ , (13)
1 − (α + Ah)
for all stepsizes h satisfying (12), where α < 1 is the positive constant in b), A is given in (7) and Q̄ h p is the bound for the local error
T n (h) defined in (5).

Hence, for Definition 1, the numerical method (2) is stable.

When the
 ∞ kernel k in (1) is of convolution
∞ type (i.e. k(t , s) = k(t − s)) assumptions b
) and c ) are equivalent respectively
∞
to ask i ) 0 k(x)dx < 1, and ii) 0 k (x)dx < +∞, thus Lemma 1 states that h j =0 ω j k j  < 1 for a sufficiently
small h, which is a classical starting point in the numerical stability analysis of convolution discrete VIEs (see for example
[8, Th. 6.20, p. 297]).
A subclass of the convolution kernels characterized by i ) and ii) consists of those satisfying in addition iii) μ(k(0)) < 0,
where μ(k) is the logarithmic norm of the matrix k (see for example [4, Sec. 7.1] for details). In this case it is more
convenient to write the error equation (4) in the explicit form
n−n0

( I − hω0k0 )en = T n (h) + h ω j k j en− j , n ≥ n0 , (14)
j =1

where again the starting errors are incorporated into T n (h). Since μ(k0 ) < 0,
1
( I − hω0 k0 )−1  ≤ < 1, (15)
1 − hω0 μ(k0 )
see [4, p. 416]. Thus, supn≥n0 en  ≤ Q̄ h p /(1 − hω0 μ(k0 ) −(α + Ah)), for all h ∈ (0, h∗ ] such that (α + Ah)/(1 − hω0 μ(k0 )) < 1,
which is a less restrictive hypothesis than (12).
A more significative improvement of the previous result, as it will be clear in Remark 2, is obtained by assuming further
iv) k(x) nonincreasing ∀x ≥ 0. In this case the following theorem holds true.
234 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237

Theorem 3. Assume that i )–iv) hold for equation (1) with convolution kernel then

Q̄ h p
sup en  ≤ , (16)
n≥n0 1 − C (h)

for all the stepsizes h such that

⎛ ⎞
1 
n0
C (h) = ⎝1 + h (ω j − ω̃ j )k0 ⎠ < 1, (17)
1 − hω0 μ(k0 )
j =1

where Q̄ h p is the bound for the local error T n (h) defined in (14) and ω̃ j , j = 0, 1, . . . , n0 are given in (8).

Proof. Assumptions i ) and iv) imply that

+∞

h ki  < 1. (18)
i =1

From (14), since (5), (18) and (15) hold,

n−n0
Q̄ h p h 
sup en  ≤ + ω j k j  sup en .
n≥n0 1 − hω0 μ(k0 ) 1 − hω0 μ(k0 ) n≥n0
j =1

Now,
⎛ ⎞
n−n0
h  1 
n0
ω j k j  ≤ ⎝1 − h (ω j − ω̃ j )k0 ⎠ .
1 − hω0 μ(k0 ) 1 − hω0 μ(k0 )
j =1 j =1

Hence, for every stepsize h satisfying (17), the bound (16) holds. 2


n0
Remark 2. Observe that, since |μ(k0 )| ≤ k0 , condition (17) is satisfied whenever ω0 > (ω j − ω̃ j ). Among the DQ–(ρ , σ )
j =1
methods the ones arising in n0 -step Gregory formulas, with n0 = 2, 3, 4 and Backward Differentiation formulas, with n0 ≤ 6,
accomplish all the sufficient conditions on the weights to be unconditionally stable.

Remark 3. From a numerical point of view, Theorem 3 extends, for example, the results in [7] to systems. Furthermore, here
an explicit expression of the bound is given.

In this section we have assumed that in (5) supn | Q (t )| ≤ Q̄ , with Q̄ independent of n and h and sup0<h<h∗ supn≥0 | R n (h)|
is finite for small h∗ > 0. This is certainly true for DQ–(ρ , σ ) methods of Gregory type where, writing down the error ex-
pression obtained in [4, sec. 2.6.3] over unbounded intervals (N → ∞), one gets | T n (h)| ≤ C 1 h + C 2 h3 , for n ≥ 0, with
C 1 and C 2 independent of h and n provided that: h1) supt ≥s k(t , s) y (s) < ∞, h2) supt ≥s  ∂∂s (k(t , s) y (s))  < ∞ and
t 3
h3) supt ≥0 0  ∂∂s3 (k(t , s) y (s)) ds < ∞. These conditions implicate some regularity on the unknown solution y (t ) of (1).
In the next section it will be clear that y (t ) is bounded for t ≥ 0, in hypotheses a) and b), when f (t ) is bounded.
By some further manipulations on equation (1) we get that hypotheses h1)–h3) are accomplished if the following as-
sumptions on the  functions involved in (1) are made:  k ∈ C 3 (Rd×d ), f ∈ C3 (Rd ), supt ≥0  f (i ) (t ) = F (i ) < ∞, i = 0, 1
 t  ∂ i k(t ,s)   ∂ k(t ,s)   t  ∂ i k(t ,s) 
supt ≥0 0  ∂ti 
ds < +∞, i = 0, 1, 2 supt ≥s  ∂ s  ds < +∞ and supt ≥0 0   ds < +∞, i = 0, 1, 2.
∂ si

4. Boundedness of the numerical solution

Since we are dealing with linear problems, all the investigations carried out on the error equation can be replicated
directly on the discrete equation (2), to obtain the following result on the boundedness of the numerical solution.

Theorem 4. Assume that a), b) and c ) hold for equation (1) and that

d) supt ≥0  f (t ) = F ,
 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237 235

Fig. 1. (a) Problem (19): error behavior of Trapezoidal DQ method for h = 1 (solid line) and h = .1 (dotted line), (b) Problem (20): numerical solution (dotted
line) vs. analytical one (solid line), when h = 0.1.

then
Fh
 yn  ≤ ,
1 − (α + Ah)
for all n ≥ 0 and h ∈ (0, h∗ ], with F h = F + n0 hW K max0≤ j ≤n0 −1 | y j | and α and A given in b) and (7) respectively.

If we assume that a), and b) hold then the solution y (t ) to the integral equation (1) is bounded, and the following result
is easily proved.

F
Theorem 5. Assume that a), b) and d) hold for equation (1), then  y (t ) ≤ , with α defined in b).
1−α

While Theorem 2 concerns the stability of the method (2) in the sense of boundedness of the global error, Theorems 4
and 5 allow comparison between the asymptotic behaviors of the analytical and numerical solution.
The bounds of the analytical and numerical solution remain at a constant distance as n → ∞ for any fixed h, and coincide
when h → 0.

5. Numerical experiments

In this section we perform numerical experiments on some mathematical problems consistent with the assumptions
made in the paper. Even if the following are theoretical examples, they have the scope to show experimentally the results
obtained in the previous sections.
We consider equation (1) with

e −s(t +1)
k(t , s) = , and f (t ), such that y (t ) = 1 + t 2 . (19)
1 + s2
The solvability of equation (2) (1 − hw nn knn = 0) does not impose significative restrictions on the stepsize h and (12) is true
independently of the steplength h. Fig. 1 (a) shows the behavior of the global error en over a long time interval for fixed
stepsizes h = 1 and h = 0.1 respectively. The method used for the integration is the Trapezoidal DQ. As a confirmation of
the theoretical result in Theorem 2, the global error is bounded by a quantity that decreases with h.
We consider equation (1) with

k(t , s) = λ2 (t − s)e −(t −s) , 0 < λ < 1,

(20)
f (t ) such that the solution is y (t ) = 14 (2λt − 1 + e −2λt ).

Equation (1)–(20) (see [4, p. 506, 16a]) is an example of linear convolution equation which satisfies i ) and ii). According
to Theorem 2 a DQ–(ρ , σ ) method of the form (2) is stable whenever the stepsize h satisfies (7). In this example we
consider the DQ method (2) whose convolution weights are generated by the fifth order Gregory DQ method. In this case,
it turns out that the restriction (12) on the stepsize is not significative (h ≤ 3.5). Furthermore, the solvability of equation
(1) (1 − hw 0 k0 = 0) is always guaranteed because k0 = 0. Fig. 1 (b) shows the plot of the analytical solution of problem
(20), with λ = 12 , in the interval [0, 100] compared with the plot of the numerical approximation obtained with h = 0.1. It
is clear that the numerical solution behaves stably. The maximum global error is 7.9 · 10−6 .
Our last experiment consists of equation (1) with

k(t , s) = t +s1+1 ,
(21)
f (t ) such that the solution is y (t ) = sin (t ).
236 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237

Fig. 2. Problem (21): error behavior of Trapezoidal (a) and fifth order Gregory (b) DQ methods for h = 1 (solid line) and h = .1 (dotted line).

Again the solvability condition (6) does not impose significative restrictions on the stepsize h and Lemma 1 is accom-
plished for each h for the Trapezoidal DQ method and for h < 1.13 in case of Gregory DQ method of order 5. Fig. 2 shows
the behavior of the global error over [0, 100] for h = 1 and h = 0.1, when Trapezoidal and order 5 Gregory DQ methods are
used, respectively.

6. Concluding remarks

We have considered systems of Volterra integral equations and analyzed the numerical stability of DQ–(ρ , σ ) methods
in terms of the boundedness of the global error. On a quite general class of problems we obtain stability results under some
conditions on the stepsize h that, in practice, turn out to be not significantly restrictive. Furthermore, in particular cases
(see assumptions iii) and iv)) the DQ methods of whatever order turn out to be unconditionally stable.
In the literature,
 t the asymptotic behavior of numerical methods is mainly analyzed with respect to the basic test equation
t
y (t ) = f (t ) + λ 0 y (s)ds, Reλ < 0 or the convolution equations y (t ) = f (t ) + 0 (λ + μ(t − s)) y (s)ds, with λ, μ < 0 and y (t ) =
t
f (t ) + λ 0 k(t − s) y (s)ds, Reλ < 0, where k(t ) ∈ L 1 (R+ ) is continuous and positive definite (see [4] and the bibliography
therein). Other papers (see for example [16] and the bibliography therein) provide results on the stability of numerical
methods with respect to general linear equations of the type (1), not necessarily of convolution type, where the results are
often based on some additional requirements on the sign of k, its derivatives and on the forcing term. The novelty of our
approach stands in the use of (1) with hypothesis b) as test problem for the numerical stability analysis of DQ methods.
As mentioned in Section 3, this problem has been treated analytically already, however no numerical studies are present
in the literature. Problems where the kernel is not positive definite  t or not of convolution type as for example the ones
in Section 5, or non-linear equations of the type y (t ) = f (t ) + 0 k(t , s) g ( y (s))ds, where ∂ g /∂ y is non-negative (see for
example the one in [15], arising in a SIS epidemic model), do not fit into the class of equations which can be treated by the
existing numerical theory and may be better represented by the analysis carried out here.

Acknowledgements

The research that led to the present paper was partially supported by a grant of the group GNCS of INdAM. The research
of the second author was partially supported by the C.N.R. MATHTECH project. The authors are grateful to the referee for
the helpful suggestions.

References

[1] M. Annunziato, H. Brunner, E. Messina, Asymptotic stability of solutions to Volterra-renewal integral equations with space maps, J. Math. Anal. Appl.
395 (2) (2012) 766–775.
[2] C. Baker, A perspective on the numerical treatment of Volterra equations, in: Numerical Analysis 2000, vol. VI, Ordinary Differential Equations and
Integral Equations J. Comput. Appl. Math. 125 (1–2) (2000) 217–249.
[3] H. Brunner, Collocation Methods for Volterra Integral and Related Functional Differential Equations, Cambridge University Press, Cambridge, UK, 2004.
[4] H. Brunner, P. van der Houwen, The Numerical Solution of Volterra Equations, North-Holland, Amsterdam, The Netherlands, 1986.
[5] T. Burton, Volterra Integral and Differential Equations, Elsevier B. V., Amsterdam, The Netherlands, 2005.
[6] C. Corduneanu, Integral Equations and Stability of Feedback Systems, Academic Press, New York/London, 1973.
[7] M. Crisci, V. Kolmanovskii, E. Russo, A. Vecchio, Boundedness of discrete Volterra equations, J. Integral Equ. Appl. 7 (4) (1995) 393–411.
[8] S. Elaydi, An Introduction to Difference Equations, Springer, New York, 2005.
[9] Z. Feng, H.Z.D. Xu, Epidemiological models with non-exponentially distributed disease stages and applications to disease control, Bull. Math. Biol. 69 (5)
(2007) 1511–1536.
[10] G. Gripenberg, S.-O. Londen, O. Staffans, Volterra Integral and Functional Equations, Cambridge University Press, Cambridge, UK, 1990.
[11] N. Hartung, Efficient resolution of metastatic tumor growth models by reformulation into integral equations, Discrete Contin. Dyn. Syst., Ser. B 20 (2)
(2015) 445–467.
[12] P. Linz, Analytical and Numerical Methods for Volterra Equations, SIAM, Philadelphia, 1985.
[13] C. Lubich, On the stability of linear multistep methods for Volterra convolution equations, IMA J. Numer. Anal. 3 (4) (1983) 439–465.
[14] E. Messina, E. Russo, A. Vecchio, Comparing analytical and numerical solution of a nonlinear two-delay integral equation, Math. Comput. Simul. 81 (5)
(2011) 1017–1026.
 E. Messina, A. Vecchio / Applied Numerical Mathematics 116 (2017) 230–237 237

[15] P. van den Driessche, J. Watmough, A simple sis epidemic model with a backward bifurcation, J. Math. Biol. 40 (6) (2010) 525–540.
[16] A. Vecchio, Volterra discrete equations: summability of the fundamental matrix, Numer. Math. 89 (4) (2001) 783–794.
[17] A. Vecchio, Stability of direct quadrature methods for systems of Volterra integral equations, J. Comput. Methods Sci. Eng. 3 (2) (2003) 71–80.
[18] P.H.M. Wolkenfelt, Linear Multistep Methods and the Construction of Quadrature Formulae for Volterra Integral and Integro-differential Equations,
Afdeling Numerieke Wiskunde (Department of Numerical Mathematics), vol. 76, Mathematisch Centrum, Amsterdam, 1979.